Optimization of Depth Map Intra Coding Algorithms Based on Machine Learning

Wenjun Song

doi:10.1109/ACCESS.2025.3525703

2025-01-01

Wenjun Song

Jiaying Yang

Qiuwen Zhang *

DOI：10.1109/ACCESS.2025.3525703

原文链接

原文求助

摘要

En 中文

As a cutting-edge field of global research, 3D video technology faces the dual challenges of large data volumes and high processing complexity. Although the most recent video coding standard VVC, surpasses HEVC in coding efficiency, dedicated research on 3D video coding remains relatively scarce. Building on existing research, this study aims to develop a 3D video coding algorithm based on VVC that lowers the complexity of the encoding procedure. We focus specifically on the depth maps in 3D video content and introduce an extreme forest model from machine learning to optimize intra-frame coding. This paper proposes a novel CU partitioning strategy implemented through a two-stage extreme forest model. First, the initial model predicts the CU partitioning type, including no partition, QuadTree partitioning, Multi-type tree horizontal partitioning, and Multi-type tree vertical partitioning. For the latter two cases, a second model further refines the partitioning into binary or ternary trees. Through this two-stage prediction mechanism, we effectively bypass CU partitioning types with low probability, significantly reducing the coding complexity. The experimental results demonstrate that the proposed algorithm saves 47.46% in encoding time while maintaining coding quality, with only a 0.26% increase in Bjontegaard Delta Bitrate. This achievement provides an effective low-complexity solution for the 3D video coding field.

Keyword:

Encoding

Prediction algorithms

Complexity theory

Video coding

Partitioning algorithms

Three-dimensional displays

Forestry

Predictive models

Adaptation models

Classification algorithms

Depth map